It even helped with developing MCPs, I could focus on writing the code rather than setup the MCP server logic and if the plugin was sandboxed - I could use community written ones. I didn't find anything that met my needs (I saw hyper-mcp but wanted something more where I could have AI generate the code from a spec and plug it into the server). So I am building jilebi. It is a MCP runtime that sandboxes plugins, manages permissions, env variables and state of the plugin so it can access only what I allow it to access. You can take a look at jilebi.ai to learn more and download it for free.
I'm looking for folks who build AI agents to try it out, let me know where it falls short of your expectations. I've written a few plugins for jilebi - you can raise a PR to this repo if you'd like to contribute. As of now, the limitations of jilebi CLI are:
* It doesn't list available plugins or show you plugins you have installed
* It doesn't support resource templates yet
* It is stuck on the 2025-03-26 MCP specification for now